Amid the global open-source war, how many AI projects have been brought back down to their true unimpressive state?
Once open source loses the support of its competitive position, the retreat speed can be faster than the expansion speed.
The global large model industry is now in a fiercely competitive landscape.
On October 20, DeepSeek released an announcement that shocked the global AI community: it fully open-sourced DeepInfer, its internally native inference framework that supports large-scale online inference services, and simultaneously released the production-grade Token codec component deepseek-recipe. The two code repositories were launched on the GitHub developer community at the same time, corresponding to Apache 2.0 and MIT licenses respectively, opening the commercially usable, modifiable and redistributable underlying infrastructure code to developers worldwide. Media commented on this move as — DeepSeek open-sourced its own "money printer", making it a real OpenAI.
Among the global top large model manufacturers, DeepSeek has thus become the first enterprise to make its native inference framework in the production environment public. Prior to this, inference scheduling, operator optimization, and Token alignment logic had long been regarded by MaaS service providers as core commercial secrets to maintain the premium of official APIs and widen the deployment cost gap with the community.
This open source move is not a simple technical gift. In the technical document "DeepSeek V4.1 Flash Large-Scale Inference Deployment Practice" released together, DeepSeek disclosed the operation baseline of its own online cluster: relying on the DeepInfer scheduling system, a single cluster carries about 155 trillion Token request traffic per day, and the peak throughput after conversion is close to 17.94 billion Tokens per second. According to the API pricing standard in off-peak hours, the revenue corresponding to each computing card per hour reaches 38 US dollars.
This system natively supports two hardware routes: NVIDIA SM100 GPU and Huawei Ascend 950 NPU. The supporting deepseek-recipe is responsible for unifying the dialogue message format, prompt encoding, streaming output parsing, tool call and JSON Schema structured return, so that the call experience after private deployment by cloud manufacturers can be consistent with DeepSeek's official API.
While the news spread, the global open source track has entered a high-density release cycle. Just two days after DeepSeek's open source move, iFlytek, another leading domestic large model enterprise, announced the official release of Spark X2.6, which achieves breakthroughs in code and agent capabilities based on fully domestic computing power for the first time, and also announced that it will continue to carry out open source operations. Six months ago, it launched AstronClaw, the first domestic enterprise-level open source and self-hosted AI skill management platform.
The pace on the overseas side is equally rapid. Meta's Llama series continues to iterate weight versions, Google continues to supplement inference optimization suites after the Gemma series of lightweight models, NVIDIA led an 800 million US dollar investment to support Reflection AI to launch the Beam large model, and completed the acquisition of Hugging Face at the same time.
It is worth mentioning that Jensen Huang took the lead in uniting 25 technology companies to release the open letter "Open Weights and U.S. AI Leadership", advocating to maintain the survival of the open source weight ecosystem and opposing one-size-fits-all regulatory bans. Microsoft, Meta and Mistral are all on this joint signature list, while Anthropic and Google's closed source team still adhere to the closed source path — the rift between the open source and closed source routes inside Silicon Valley has been publicly placed on the table.
The unavoidable question then arises: Why have global large model enterprises collectively turned to open source in just two years? If open source means that model weights, inference tools and underlying codes are open to the public, which means anyone can download, fine-tune, deploy and transform them, is the moat of traditional API sales disappearing?
01
To understand this wave of AI open source, the most appropriate reference sample is not the current large model, but two classic games that have been written into the history of the software industry: Microsoft's decades-long strategy of swinging repeatedly between closure and open source, and Google's commercial experiment that Android quickly occupied the smartphone market relying on open source strategy.
The entanglement between Microsoft and open source is the best historical negative to understand this competition. From the end of the 20th century to the beginning of the 21st century, Microsoft regarded Linux and GPL as a fundamental threat to the Windows authorization model. Then CEO Steve Ballmer once called Linux a "cancer". But after Satya Nadella took over as CEO in 2014, the trend reversed.
In June 2018, Microsoft acquired GitHub for 7.5 billion US dollars. Nadella's wording in the official statement was: "Microsoft is a developer-first company. By joining GitHub, we have strengthened our commitment to developer freedom, openness and innovation." He also said the sentence that was repeatedly quoted later: "Microsoft is all-in on open source." The reason why Microsoft embraced open source as if it had "returned to its original form" is not that open source suddenly became morally correct, but that its profit center shifted from Windows authorization to Azure cloud services, and a large number of Linux and open source workloads run on the cloud. Without supporting open source, Azure would only be a "Windows Only" cloud, which has no chance of winning in the cloud war.
The Android system built by Google in the mobile Internet era verified the power of open source expansion even earlier. After the release of iPhone in 2007, Google had almost no foothold in the mobile operating system. If Android was closed source, no mobile phone manufacturer would have reason to abandon the system they were using to bet on a newcomer. Google's choice was to open source the underlying code of Android, attract Samsung, HTC, Motorola and other manufacturers to enter the market collectively with zero authorization cost, and fight against iOS by the way of "uniting all manufacturers".
But Google's "open source" was designed in layers from the very beginning: the underlying code was opened to expand the device base, and GMS — Play Store, Gmail, Maps — which truly generates continuous profits, has always been closed source and firmly controlled by Google. The victory of Android is the victory of the business architecture of making scale through the open layer and making profits through the closed source layer, rather than the victory of the pure open source concept.
In short, the underlying logic of Android is to open the underlying layer to compete for ecological seats, and close the traffic entrance to hold the monetization channel. Applied to today's large model industry, DeepSeek's open source of DeepInfer this time is essentially an upgraded version that reproduces this logic: the model weights have been opened before, and now it further opens the inference scheduling middle layer, fills the most expensive and highest threshold link in the deployment link, reduces the cost of cloud manufacturers, government and enterprise customers to build their own services, while DeepSeek still retains high-value businesses such as natively optimized upper-layer services, enterprise privatization consulting, and customized fine-tuning. It no longer only earns the price difference by selling Token APIs, but turns to expand the commercial contact surface of the entire technology stack relying on the ecological scale.
Putting these two histories together, we can get a simple but critical judgment: when enterprises choose open source, it is usually not because open source itself is the goal, but because in a specific competitive position, it is the most effective lever to expand the ecosystem, share costs, and fight against incumbent monopolists. When this competitive position changes, the boundary and sincerity of open source will also change accordingly.
Moreover, the changes at the geopolitical level have further amplified the strategic value of open source. Domestic AI manufacturers are facing the external environment of limited supply of high-end chips, and they do not have natural advantages when simply competing for the training computing power of super-large clusters. Open source has become a differentiated breakthrough path: release models, operators, and inference frameworks to the global community, rely on the optimization of the algorithm layer and scheduling layer to build competitive models under limited computing power conditions, and use the developer community to spontaneously complete adaptation, fine-tuning, and scenario implementation, so that millions of developers around the world can jointly complete ecological construction, without the need for enterprises to invest huge market budgets alone to educate customers.
As for NVIDIA's support for open source, its essence is to prevent the global open source ecosystem from accelerating away from NVIDIA's hardware system and damaging its own hardware business. In this game, open source is no longer just a product strategy, but also a geopolitical chip for computing power manufacturers and model manufacturers to compete for ecological standards.
02
However, under the boom, a cruel question hangs over all open source players: open source does not equal victory. Opening weights and underlying codes only means getting the admission ticket to the ecological competition field. A large number of open source models that were once very popular quickly fell silent after the heat faded, that is, as the title says, they returned to their original form in the fierce competition.
In fact, the past open source software field has repeatedly staged such plots: countless open source projects became popular briefly, then stagnated slowly after the core team's resources were exhausted, and were eventually forgotten by the ecosystem. In the large model track, the elimination speed will only be faster. The model evaluation standards are continuously updated, the inference hardware is continuously iterated, and customers' requirements for stability, multimodality and agent capabilities are rising. Relying only on one-time open source cannot hold the position.
Deducted along the current industrial clues, the future evolution path of this global open source war will not lead to a single winner taking all, but to a pattern of stratification, division of camps, and deep binding of software and hardware. Closed source and open source will not eliminate each other, but coexist and penetrate each other for a long time.
From the current global large model market, the first layer of the track is the competition of underlying inference and deployment infrastructure, which is the battlefield of inference engines, operator libraries, and codec components. At this layer, compatibility, hardware adaptation, throughput latency, and cost bottom line are the core indicators. Its profit method is not to sell code directly, but to support enterprise-level support, cluster deployment consulting, customized operator development, and computing power scheduling services.
The winner at this layer will not be a single manufacturer. A few frameworks that have been verified by trillion-level Token traffic and are compatible with multiple hardware architectures will form de facto standards. DeepSeek's bet on DeepInfer may be to enter this camp and turn its engineering experience in large-scale inference into industry standards.
The second layer is the competition of base model weights, which is also the track with the highest public attention. There will be an obvious differentiation route for base models: some models take the general all-round route, continuously benchmarking the capabilities of closed source flagship, taking into account code, mathematics, long text and multimodality; the other part of models will most likely move towards vertical lightweight, targeting end-side, industry, medical, vehicle-mounted and other subdivided scenarios, pursuing low power consumption, low cost and high certainty.
Please note that long-term stable base model players must meet two conditions at the same time: continuous and stable computing power investment to support iteration, and a large and active developer community that continuously outputs fine-tuned versions and application plugins. The large models with high domestic activity are all concentrated at this level, but everyone can count the players that can meet both conditions at the same time.
The third layer is the binding camp of hardware-model-inference stack, which is the battlefield that NVIDIA cares most about, and it is also the core opportunity for domestic open source manufacturers. The stronger the heterogeneous adaptation capability of open source models, the faster the ecological short board of domestic computing power can be made up.
Conversely, the maturity of the computing power base also determines whether open source models can be deployed on a large scale in government and enterprises. In the future, when customers choose open source models, most of the time they will not look at the evaluation scores first, but see if it can run stably on their existing computing power clusters. The power of the hardware ecosystem will continue to reshape the market share of models. From this perspective, iFlytek is the only domestic large model manufacturer with a domestic computing power base, and may have more say in the future.
Fairly speaking, for Chinese AI enterprises, the open source wave brings obvious dividends, and at the same time it also buries structural problems that are difficult to avoid.
The most obvious manifestation of the dividend is the ecological discourse power. Relying on open source models, domestic manufacturers bypass the high-end computing power barrier, obtain channels to reach global developers, government and enterprise customers, establish brand awareness in markets that attach importance to data autonomy such as Europe, Southeast Asia, and the Middle East, and break the pattern that AI capabilities can only be supplied by a few American closed source manufacturers in the past.
Open source also forces the evolution of internal engineering capabilities. Community testing, vulnerability feedback, and secondary development are equivalent to free global quality verification. After the inference framework and Token components are opened, developers around the world will conduct pressure tests on extreme scenarios on different hardware, helping manufacturers find defects that internal tests cannot cover. At the same time, open source also creates an ecological soil for domestic computing power. For NPU platforms such as Ascend to grow, what they need most is the active adaptation of a large number of models, operators, and inference frameworks.
But the challenges cannot be underestimated at all.
The core problem is the difficulty of commercial closed loop. After open source, the model weight itself is no longer a source of revenue. Enterprises must turn to monetization through private deployment, technical support, fine-tuning services, computing power leasing, and industry solutions. Such businesses are heavy delivery and long-cycle projects, with gross profit much higher than API calls, but it is more difficult to scale up. Many open source model teams have good technology, but lack a delivery network for government and enterprises, so the community popularity is difficult to convert into stable cash flow. Once the financing environment tightens, the funds for continuous iteration will be broken, and the project will stagnate, which is the so-called "return to original form".
Governance and compliance risks are unavoidable topics. After open source codes and model weights are distributed externally, the model output content, data traceability, copyright, distillation derivation, export control, and local regulatory rules become complicated. Different countries have huge differences in the control standards for open source weights. Once the model flows out, it is difficult to control the scenarios of downstream users, and the compliance cost continues to rise.
What is easy to ignore is the long-term cost of community governance. The open source community cannot operate automatically just by releasing codes. It requires continuous maintenance of documents, merging community contributions, maintaining version branches, coordinating derivative projects, etc., which requires a dedicated engineering team to invest for a long time. Many domestic teams underestimate the human consumption of community operation.
Homogenized involution is also a big hurdle. A large number of open source models have close indicators on the evaluation list, with insufficient differentiation. Customers are prone to fall into price comparison when choosing, which compresses the profit space of all players. The industry is also easy to fall into vicious price wars, which has occurred many times in China's photovoltaic, home appliance and other industries.
Back to the original topic, who will stay at the poker table?
The global open source war will not end with "open source defeating closed source" or "closed source eliminating open source". The more likely pattern is that the boundary of open source is re-divided: Weight open source becomes the norm, inference framework open source becomes a new battlefield for a few players with real engineering depth, while closed source capabilities retreat to the cutting-edge training that requires continuous large-scale computing power investment and the highly regulated enterprise service layer.
In this pattern, the one that can "laugh last" will not be the one that shouts the open source slogan the loudest, but the type of participant that can establish a structural connection between open source and business. NVIDIA's bet is to use capital and platforms to keep open source developers on its own hardware track; Google uses Gemma to maintain its presence on the open source side, while using Gemini to hold the closed source cutting edge. But Meta's exit reminds all participants that Once open source loses the support of its competitive position, the retreat speed can be faster than the expansion speed.
As for how many AI players will return to their original form in the competition, the answer will definitely not come from any announcement, but from the frequency of submission records of those open source projects in the next few quarters and years, the question quality of enterprise users in the issue area, and the next round of price reconstruction centered on inference cost.
【This article is for communication only, not investment advice. Please pay attention to investment risks.】
This article is from the WeChat official account "Daily Capital Theory", Author: Reflections of Plain Boiled Water, Published with authorization from 36Kr.